intel: Share URB configuration code between GL and Vulkan.

This code is far too complicated to cut and paste.

v2: Update the newly added genX_gpu_memcpy.c; const a few things.

Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Topi Pohjolainen <[email protected]>

7 files changed, 252 insertions, 156 deletions

diff --git a/src/intel/Makefile.sources b/src/intel/Makefile.sources
index c1740fe3b57..be6bdef7b69 100644
--- a/src/intel/Makefile.sources
+++ b/src/intel/Makefile.sources
@@ -11,6 +11,7 @@ COMMON_FILES = \
 	common/gen_device_info.h \
 	common/gen_l3_config.c \
 	common/gen_l3_config.h \
+	common/gen_urb_config.c \
 	common/gen_sample_positions.h
 
 GENXML_GENERATED_FILES = \
diff --git a/src/intel/common/gen_l3_config.h b/src/intel/common/gen_l3_config.h
index 25a9675f462..8dc7dda0fcc 100644
--- a/src/intel/common/gen_l3_config.h
+++ b/src/intel/common/gen_l3_config.h
@@ -92,4 +92,10 @@ gen_get_l3_config_urb_size(const struct gen_device_info *devinfo,
 
 void gen_dump_l3_config(const struct gen_l3_config *cfg, FILE *fp);
 
+void gen_get_urb_config(const struct gen_device_info *devinfo,
+                        unsigned push_constant_bytes, unsigned urb_size_bytes,
+                        bool tess_present, bool gs_present,
+                        const unsigned entry_size[4],
+                        unsigned entries[4], unsigned start[4]);
+
 #endif /* GEN_L3_CONFIG_H */
diff --git a/src/intel/common/gen_urb_config.c b/src/intel/common/gen_urb_config.c
new file mode 100644
index 00000000000..c925a0dbb36
--- /dev/null
+++ b/src/intel/common/gen_urb_config.c
@@ -0,0 +1,201 @@
+/*
+ * Copyright (c) 2011 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <stdlib.h>
+#include <math.h>
+
+#include "util/macros.h"
+#include "main/macros.h"
+
+#include "gen_l3_config.h"
+
+/**
+ * The following diagram shows how we partition the URB:
+ *
+ *        16kb or 32kb               Rest of the URB space
+ *   __________-__________   _________________-_________________
+ *  /                     \ /                                   \
+ * +-------------------------------------------------------------+
+ * |  VS/HS/DS/GS/FS Push  |           VS/HS/DS/GS URB           |
+ * |       Constants       |               Entries               |
+ * +-------------------------------------------------------------+
+ *
+ * Push constants must be stored at the beginning of the URB space,
+ * while URB entries can be stored anywhere.  We choose to lay them
+ * out in pipeline order (VS -> HS -> DS -> GS).
+ */
+
+/**
+ * Decide how to partition the URB among the various stages.
+ *
+ * \param[in] push_constant_bytes - space allocate for push constants.
+ * \param[in] urb_size_bytes - total size of the URB (from L3 config).
+ * \param[in] tess_present - are tessellation shaders active?
+ * \param[in] gs_present - are geometry shaders active?
+ * \param[in] entry_size - the URB entry size (from the shader compiler)
+ * \param[out] entries - the number of URB entries for each stage
+ * \param[out] start - the starting offset for each stage
+ */
+void
+gen_get_urb_config(const struct gen_device_info *devinfo,
+                   unsigned push_constant_bytes, unsigned urb_size_bytes,
+                   bool tess_present, bool gs_present,
+                   const unsigned entry_size[4],
+                   unsigned entries[4], unsigned start[4])
+{
+   const bool active[4] = { true, tess_present, tess_present, gs_present };
+
+   /* URB allocations must be done in 8k chunks. */
+   const unsigned chunk_size_bytes = 8192;
+
+   const unsigned push_constant_chunks =
+      push_constant_bytes / chunk_size_bytes;
+   const unsigned urb_chunks = urb_size_bytes / chunk_size_bytes;
+
+   /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
+    *
+    *     VS Number of URB Entries must be divisible by 8 if the VS URB Entry
+    *     Allocation Size is less than 9 512-bit URB entries.
+    *
+    * Similar text exists for HS, DS and GS.
+    */
+   unsigned granularity[4];
+   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+      granularity[i] = (entry_size[i] < 9) ? 8 : 1;
+   }
+
+   unsigned min_entries[4] = {
+      /* VS has a lower limit on the number of URB entries.
+       *
+       * From the Broadwell PRM, 3DSTATE_URB_VS instruction:
+       * "When tessellation is enabled, the VS Number of URB Entries must be
+       *  greater than or equal to 192."
+       */
+      [MESA_SHADER_VERTEX] = tess_present && devinfo->gen == 8 ?
+         192 : devinfo->urb.min_entries[MESA_SHADER_VERTEX],
+
+      /* There are two constraints on the minimum amount of URB space we can
+       * allocate:
+       *
+       * (1) We need room for at least 2 URB entries, since we always operate
+       * the GS in DUAL_OBJECT mode.
+       *
+       * (2) We can't allocate less than nr_gs_entries_granularity.
+       */
+      [MESA_SHADER_GEOMETRY] = gs_present ? 2 : 0,
+
+      [MESA_SHADER_TESS_CTRL] = tess_present ? 1 : 0,
+
+      [MESA_SHADER_TESS_EVAL] = tess_present ?
+         devinfo->urb.min_entries[MESA_SHADER_TESS_EVAL] : 0,
+   };
+
+   /* Min VS Entries isn't a multiple of 8 on Cherryview/Broxton; round up.
+    * Round them all up.
+    */
+   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+      min_entries[i] = ALIGN(min_entries[i], granularity[i]);
+   }
+
+   unsigned entry_size_bytes[4];
+   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+      entry_size_bytes[i] = 64 * entry_size[i];
+   }
+
+   /* Initially, assign each stage the minimum amount of URB space it needs,
+    * and make a note of how much additional space it "wants" (the amount of
+    * additional space it could actually make use of).
+    */
+   unsigned chunks[4];
+   unsigned wants[4];
+   unsigned total_needs = push_constant_chunks;
+   unsigned total_wants = 0;
+
+   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+      if (active[i]) {
+         chunks[i] = DIV_ROUND_UP(min_entries[i] * entry_size_bytes[i],
+                                  chunk_size_bytes);
+
+         wants[i] =
+            DIV_ROUND_UP(devinfo->urb.max_entries[i] * entry_size_bytes[i],
+                         chunk_size_bytes) - chunks[i];
+      } else {
+         chunks[i] = 0;
+         wants[i] = 0;
+      }
+
+      total_needs += chunks[i];
+      total_wants += wants[i];
+   }
+
+   assert(total_needs <= urb_chunks);
+
+   /* Mete out remaining space (if any) in proportion to "wants". */
+   unsigned remaining_space = MIN2(urb_chunks - total_needs, total_wants);
+
+   if (remaining_space > 0) {
+      for (int i = MESA_SHADER_VERTEX;
+           total_wants > 0 && i <= MESA_SHADER_TESS_EVAL; i++) {
+         unsigned additional = (unsigned)
+            roundf(wants[i] * (((float) remaining_space) / total_wants));
+         chunks[i] += additional;
+         remaining_space -= additional;
+         total_wants -= additional;
+      }
+
+      chunks[MESA_SHADER_GEOMETRY] += remaining_space;
+   }
+
+   /* Sanity check that we haven't over-allocated. */
+   unsigned total_chunks = push_constant_chunks;
+   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+      total_chunks += chunks[i];
+   }
+   assert(total_chunks <= urb_chunks);
+
+   /* Finally, compute the number of entries that can fit in the space
+    * allocated to each stage.
+    */
+   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+      entries[i] = chunks[i] * chunk_size_bytes / entry_size_bytes[i];
+
+      /* Since we rounded up when computing wants[], this may be slightly
+       * more than the maximum allowed amount, so correct for that.
+       */
+      entries[i] = MIN2(entries[i], devinfo->urb.max_entries[i]);
+
+      /* Ensure that we program a multiple of the granularity. */
+      entries[i] = ROUND_DOWN_TO(entries[i], granularity[i]);
+
+      /* Finally, sanity check to make sure we have at least the minimum
+       * number of entries needed for each stage.
+       */
+      assert(entries[i] >= min_entries[i]);
+   }
+
+   /* Lay out the URB in pipeline order: push constants, VS, HS, DS, GS. */
+   start[0] = push_constant_chunks;
+   for (int i = MESA_SHADER_TESS_CTRL; i <= MESA_SHADER_GEOMETRY; i++) {
+      start[i] = start[i - 1] + chunks[i - 1];
+   }
+}
diff --git a/src/intel/vulkan/anv_genX.h b/src/intel/vulkan/anv_genX.h
index 793717012b6..44868936eb2 100644
--- a/src/intel/vulkan/anv_genX.h
+++ b/src/intel/vulkan/anv_genX.h
@@ -55,9 +55,9 @@ void genX(cmd_buffer_flush_compute_state)(struct anv_cmd_buffer *cmd_buffer);
 
 void
 genX(emit_urb_setup)(struct anv_device *device, struct anv_batch *batch,
+                     const struct gen_l3_config *l3_config,
                      VkShaderStageFlags active_stages,
-                     unsigned vs_entry_size, unsigned gs_entry_size,
-                     const struct gen_l3_config *l3_config);
+                     const unsigned entry_size[4]);
 
 void genX(cmd_buffer_emit_hz_op)(struct anv_cmd_buffer *cmd_buffer,
                                enum blorp_hiz_op op);
diff --git a/src/intel/vulkan/genX_blorp_exec.c b/src/intel/vulkan/genX_blorp_exec.c
index a705de0d345..a07f370a14e 100644
--- a/src/intel/vulkan/genX_blorp_exec.c
+++ b/src/intel/vulkan/genX_blorp_exec.c
@@ -124,11 +124,13 @@ blorp_emit_urb_config(struct blorp_batch *batch, unsigned vs_entry_size)
    struct anv_device *device = batch->blorp->driver_ctx;
    struct anv_cmd_buffer *cmd_buffer = batch->driver_batch;
 
+   const unsigned entry_size[4] = { vs_entry_size, 1, 1, 1 };
+
    genX(emit_urb_setup)(device, &cmd_buffer->batch,
+                        cmd_buffer->state.current_l3_config,
                         VK_SHADER_STAGE_VERTEX_BIT |
                         VK_SHADER_STAGE_FRAGMENT_BIT,
-                        vs_entry_size, 0,
-                        cmd_buffer->state.current_l3_config);
+                        entry_size);
 }
 
 void genX(blorp_exec)(struct blorp_batch *batch,
diff --git a/src/intel/vulkan/genX_gpu_memcpy.c b/src/intel/vulkan/genX_gpu_memcpy.c
index 10751b0a35e..eb11c2f0874 100644
--- a/src/intel/vulkan/genX_gpu_memcpy.c
+++ b/src/intel/vulkan/genX_gpu_memcpy.c
@@ -149,9 +149,11 @@ genX(cmd_buffer_gpu_memcpy)(struct anv_cmd_buffer *cmd_buffer,
     * allocate space for the VS.  Even though one isn't run, we need VUEs to
     * store the data that VF is going to pass to SOL.
     */
+   const unsigned entry_size[4] = { DIV_ROUND_UP(32, 64), 1, 1, 1 };
+
    genX(emit_urb_setup)(cmd_buffer->device, &cmd_buffer->batch,
-                        VK_SHADER_STAGE_VERTEX_BIT, DIV_ROUND_UP(32, 64), 0,
-                        cmd_buffer->state.current_l3_config);
+                        cmd_buffer->state.current_l3_config,
+                        VK_SHADER_STAGE_VERTEX_BIT, entry_size);
 
    anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_SO_BUFFER), sob) {
       sob.SOBufferIndex = 0;
diff --git a/src/intel/vulkan/genX_pipeline.c b/src/intel/vulkan/genX_pipeline.c
index 89c9caf9fea..c54265c656d 100644
--- a/src/intel/vulkan/genX_pipeline.c
+++ b/src/intel/vulkan/genX_pipeline.c
@@ -183,129 +183,27 @@ emit_vertex_input(struct anv_pipeline *pipeline,
 
 void
 genX(emit_urb_setup)(struct anv_device *device, struct anv_batch *batch,
+                     const struct gen_l3_config *l3_config,
                      VkShaderStageFlags active_stages,
-                     unsigned vs_size, unsigned gs_size,
-                     const struct gen_l3_config *l3_config)
+                     const unsigned entry_size[4])
 {
-   if (!(active_stages & VK_SHADER_STAGE_VERTEX_BIT))
-      vs_size = 1;
-
-   if (!(active_stages & VK_SHADER_STAGE_GEOMETRY_BIT))
-      gs_size = 1;
-
-   unsigned vs_entry_size_bytes = vs_size * 64;
-   unsigned gs_entry_size_bytes = gs_size * 64;
-
-   /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
-    *
-    *     VS Number of URB Entries must be divisible by 8 if the VS URB Entry
-    *     Allocation Size is less than 9 512-bit URB entries.
-    *
-    * Similar text exists for GS.
-    */
-   unsigned vs_granularity = (vs_size < 9) ? 8 : 1;
-   unsigned gs_granularity = (gs_size < 9) ? 8 : 1;
-
-   /* URB allocations must be done in 8k chunks. */
-   unsigned chunk_size_bytes = 8192;
-
-   /* Determine the size of the URB in chunks. */
-   const unsigned total_urb_size =
-      gen_get_l3_config_urb_size(&device->info, l3_config);
-   const unsigned urb_chunks = total_urb_size * 1024 / chunk_size_bytes;
-
-   /* Reserve space for push constants */
-   unsigned push_constant_kb;
-   if (device->info.gen >= 8)
-      push_constant_kb = 32;
-   else if (device->info.is_haswell)
-      push_constant_kb = device->info.gt == 3 ? 32 : 16;
-   else
-      push_constant_kb = 16;
-
-   unsigned push_constant_bytes = push_constant_kb * 1024;
-   unsigned push_constant_chunks =
-      push_constant_bytes / chunk_size_bytes;
-
-   /* Initially, assign each stage the minimum amount of URB space it needs,
-    * and make a note of how much additional space it "wants" (the amount of
-    * additional space it could actually make use of).
-    */
-
-   /* VS has a lower limit on the number of URB entries */
-   unsigned vs_chunks =
-      ALIGN(device->info.urb.min_entries[MESA_SHADER_VERTEX] *
-            vs_entry_size_bytes, chunk_size_bytes) / chunk_size_bytes;
-   unsigned vs_wants =
-      ALIGN(device->info.urb.max_entries[MESA_SHADER_VERTEX] *
-            vs_entry_size_bytes,
-            chunk_size_bytes) / chunk_size_bytes - vs_chunks;
-
-   unsigned gs_chunks = 0;
-   unsigned gs_wants = 0;
-   if (active_stages & VK_SHADER_STAGE_GEOMETRY_BIT) {
-      /* There are two constraints on the minimum amount of URB space we can
-       * allocate:
-       *
-       * (1) We need room for at least 2 URB entries, since we always operate
-       * the GS in DUAL_OBJECT mode.
-       *
-       * (2) We can't allocate less than nr_gs_entries_granularity.
-       */
-      gs_chunks = ALIGN(MAX2(gs_granularity, 2) * gs_entry_size_bytes,
-                        chunk_size_bytes) / chunk_size_bytes;
-      gs_wants =
-         ALIGN(device->info.urb.max_entries[MESA_SHADER_GEOMETRY] *
-               gs_entry_size_bytes,
-               chunk_size_bytes) / chunk_size_bytes - gs_chunks;
-   }
-
-   /* There should always be enough URB space to satisfy the minimum
-    * requirements of each stage.
-    */
-   unsigned total_needs = push_constant_chunks + vs_chunks + gs_chunks;
-   assert(total_needs <= urb_chunks);
-
-   /* Mete out remaining space (if any) in proportion to "wants". */
-   unsigned total_wants = vs_wants + gs_wants;
-   unsigned remaining_space = urb_chunks - total_needs;
-   if (remaining_space > total_wants)
-      remaining_space = total_wants;
-   if (remaining_space > 0) {
-      unsigned vs_additional = (unsigned)
-         round(vs_wants * (((double) remaining_space) / total_wants));
-      vs_chunks += vs_additional;
-      remaining_space -= vs_additional;
-      gs_chunks += remaining_space;
-   }
-
-   /* Sanity check that we haven't over-allocated. */
-   assert(push_constant_chunks + vs_chunks + gs_chunks <= urb_chunks);
-
-   /* Finally, compute the number of entries that can fit in the space
-    * allocated to each stage.
-    */
-   unsigned nr_vs_entries = vs_chunks * chunk_size_bytes / vs_entry_size_bytes;
-   unsigned nr_gs_entries = gs_chunks * chunk_size_bytes / gs_entry_size_bytes;
-
-   /* Since we rounded up when computing *_wants, this may be slightly more
-    * than the maximum allowed amount, so correct for that.
-    */
-   nr_vs_entries = MIN2(nr_vs_entries,
-                        device->info.urb.max_entries[MESA_SHADER_VERTEX]);
-   nr_gs_entries = MIN2(nr_gs_entries,
-                        device->info.urb.max_entries[MESA_SHADER_GEOMETRY]);
+   const struct gen_device_info *devinfo = &device->info;
+#if GEN_IS_HASWELL
+   const unsigned push_constant_kb = devinfo->gt == 3 ? 32 : 16;
+#else
+   const unsigned push_constant_kb = GEN_GEN >= 8 ? 32 : 16;
+#endif
 
-   /* Ensure that we program a multiple of the granularity. */
-   nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, vs_granularity);
-   nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, gs_granularity);
+   const unsigned urb_size_kb = gen_get_l3_config_urb_size(devinfo, l3_config);
 
-   /* Finally, sanity check to make sure we have at least the minimum number
-    * of entries needed for each stage.
-    */
-   assert(nr_vs_entries >= device->info.urb.min_entries[MESA_SHADER_VERTEX]);
-   if (active_stages & VK_SHADER_STAGE_GEOMETRY_BIT)
-      assert(nr_gs_entries >= 2);
+   unsigned entries[4];
+   unsigned start[4];
+   gen_get_urb_config(devinfo,
+                      1024 * push_constant_kb, 1024 * urb_size_kb,
+                      active_stages &
+                         VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT,
+                      active_stages & VK_SHADER_STAGE_GEOMETRY_BIT,
+                      entry_size, entries, start);
 
 #if GEN_GEN == 7 && !GEN_IS_HASWELL
    /* From the IVB PRM Vol. 2, Part 1, Section 3.2.1:
@@ -323,45 +221,31 @@ genX(emit_urb_setup)(struct anv_device *device, struct anv_batch *batch,
    }
 #endif
 
-   /* Lay out the URB in the following order:
-    * - push constants
-    * - VS
-    * - GS
-    */
-   anv_batch_emit(batch, GENX(3DSTATE_URB_VS), urb) {
-      urb.VSURBStartingAddress      = push_constant_chunks;
-      urb.VSURBEntryAllocationSize  = vs_size - 1;
-      urb.VSNumberofURBEntries      = nr_vs_entries;
-   }
-
-   anv_batch_emit(batch, GENX(3DSTATE_URB_HS), urb) {
-      urb.HSURBStartingAddress      = push_constant_chunks;
-   }
-
-   anv_batch_emit(batch, GENX(3DSTATE_URB_DS), urb) {
-      urb.DSURBStartingAddress      = push_constant_chunks;
-   }
-
-   anv_batch_emit(batch, GENX(3DSTATE_URB_GS), urb) {
-      urb.GSURBStartingAddress      = push_constant_chunks + vs_chunks;
-      urb.GSURBEntryAllocationSize  = gs_size - 1;
-      urb.GSNumberofURBEntries      = nr_gs_entries;
+   for (int i = 0; i <= MESA_SHADER_GEOMETRY; i++) {
+      anv_batch_emit(batch, GENX(3DSTATE_URB_VS), urb) {
+         urb._3DCommandSubOpcode      += i;
+         urb.VSURBStartingAddress      = start[i];
+         urb.VSURBEntryAllocationSize  = entry_size[i] - 1;
+         urb.VSNumberofURBEntries      = entries[i];
+      }
    }
 }
 
 static inline void
 emit_urb_setup(struct anv_pipeline *pipeline)
 {
-   unsigned vs_entry_size =
-      (pipeline->active_stages & VK_SHADER_STAGE_VERTEX_BIT) ?
-      get_vs_prog_data(pipeline)->base.urb_entry_size : 0;
-   unsigned gs_entry_size =
-      (pipeline->active_stages & VK_SHADER_STAGE_GEOMETRY_BIT) ?
-      get_gs_prog_data(pipeline)->base.urb_entry_size : 0;
+   unsigned entry_size[4];
+   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+      const struct brw_vue_prog_data *prog_data =
+         !anv_pipeline_has_stage(pipeline, i) ? NULL :
+         (const struct brw_vue_prog_data *) pipeline->shaders[i]->prog_data;
+
+      entry_size[i] = prog_data ? prog_data->urb_entry_size : 1;
+   }
 
    genX(emit_urb_setup)(pipeline->device, &pipeline->batch,
-                        pipeline->active_stages, vs_entry_size, gs_entry_size,
-                        pipeline->urb.l3_config);
+                        pipeline->urb.l3_config,
+                        pipeline->active_stages, entry_size);
 }
 
 static void